What would be the point running a serial program on a GPU? The problem isn't cuda or opencl. The problem is that either the math isn't parallelizable or people aren't smart enough (yet) to make it so.